archive: stop passing "stage" through read_tree_recursive()
[git] / packfile.c
1 #include "cache.h"
2 #include "list.h"
3 #include "pack.h"
4 #include "repository.h"
5 #include "dir.h"
6 #include "mergesort.h"
7 #include "packfile.h"
8 #include "delta.h"
9 #include "streaming.h"
10 #include "hash-lookup.h"
11 #include "commit.h"
12 #include "object.h"
13 #include "tag.h"
14 #include "tree-walk.h"
15 #include "tree.h"
16 #include "object-store.h"
17 #include "midx.h"
18 #include "commit-graph.h"
19 #include "promisor-remote.h"
20
21 char *odb_pack_name(struct strbuf *buf,
22                     const unsigned char *hash,
23                     const char *ext)
24 {
25         strbuf_reset(buf);
26         strbuf_addf(buf, "%s/pack/pack-%s.%s", get_object_directory(),
27                     hash_to_hex(hash), ext);
28         return buf->buf;
29 }
30
31 char *sha1_pack_name(const unsigned char *sha1)
32 {
33         static struct strbuf buf = STRBUF_INIT;
34         return odb_pack_name(&buf, sha1, "pack");
35 }
36
37 char *sha1_pack_index_name(const unsigned char *sha1)
38 {
39         static struct strbuf buf = STRBUF_INIT;
40         return odb_pack_name(&buf, sha1, "idx");
41 }
42
43 static unsigned int pack_used_ctr;
44 static unsigned int pack_mmap_calls;
45 static unsigned int peak_pack_open_windows;
46 static unsigned int pack_open_windows;
47 static unsigned int pack_open_fds;
48 static unsigned int pack_max_fds;
49 static size_t peak_pack_mapped;
50 static size_t pack_mapped;
51
52 #define SZ_FMT PRIuMAX
53 static inline uintmax_t sz_fmt(size_t s) { return s; }
54
55 void pack_report(void)
56 {
57         fprintf(stderr,
58                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
59                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
60                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
61                 sz_fmt(getpagesize()),
62                 sz_fmt(packed_git_window_size),
63                 sz_fmt(packed_git_limit));
64         fprintf(stderr,
65                 "pack_report: pack_used_ctr            = %10u\n"
66                 "pack_report: pack_mmap_calls          = %10u\n"
67                 "pack_report: pack_open_windows        = %10u / %10u\n"
68                 "pack_report: pack_mapped              = "
69                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
70                 pack_used_ctr,
71                 pack_mmap_calls,
72                 pack_open_windows, peak_pack_open_windows,
73                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
74 }
75
76 /*
77  * Open and mmap the index file at path, perform a couple of
78  * consistency checks, then record its information to p.  Return 0 on
79  * success.
80  */
81 static int check_packed_git_idx(const char *path, struct packed_git *p)
82 {
83         void *idx_map;
84         size_t idx_size;
85         int fd = git_open(path), ret;
86         struct stat st;
87         const unsigned int hashsz = the_hash_algo->rawsz;
88
89         if (fd < 0)
90                 return -1;
91         if (fstat(fd, &st)) {
92                 close(fd);
93                 return -1;
94         }
95         idx_size = xsize_t(st.st_size);
96         if (idx_size < 4 * 256 + hashsz + hashsz) {
97                 close(fd);
98                 return error("index file %s is too small", path);
99         }
100         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
101         close(fd);
102
103         ret = load_idx(path, hashsz, idx_map, idx_size, p);
104
105         if (ret)
106                 munmap(idx_map, idx_size);
107
108         return ret;
109 }
110
111 int load_idx(const char *path, const unsigned int hashsz, void *idx_map,
112              size_t idx_size, struct packed_git *p)
113 {
114         struct pack_idx_header *hdr = idx_map;
115         uint32_t version, nr, i, *index;
116
117         if (idx_size < 4 * 256 + hashsz + hashsz)
118                 return error("index file %s is too small", path);
119         if (idx_map == NULL)
120                 return error("empty data");
121
122         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
123                 version = ntohl(hdr->idx_version);
124                 if (version < 2 || version > 2)
125                         return error("index file %s is version %"PRIu32
126                                      " and is not supported by this binary"
127                                      " (try upgrading GIT to a newer version)",
128                                      path, version);
129         } else
130                 version = 1;
131
132         nr = 0;
133         index = idx_map;
134         if (version > 1)
135                 index += 2;  /* skip index header */
136         for (i = 0; i < 256; i++) {
137                 uint32_t n = ntohl(index[i]);
138                 if (n < nr)
139                         return error("non-monotonic index %s", path);
140                 nr = n;
141         }
142
143         if (version == 1) {
144                 /*
145                  * Total size:
146                  *  - 256 index entries 4 bytes each
147                  *  - 24-byte entries * nr (object ID + 4-byte offset)
148                  *  - hash of the packfile
149                  *  - file checksum
150                  */
151                 if (idx_size != st_add(4 * 256 + hashsz + hashsz, st_mult(nr, hashsz + 4)))
152                         return error("wrong index v1 file size in %s", path);
153         } else if (version == 2) {
154                 /*
155                  * Minimum size:
156                  *  - 8 bytes of header
157                  *  - 256 index entries 4 bytes each
158                  *  - object ID entry * nr
159                  *  - 4-byte crc entry * nr
160                  *  - 4-byte offset entry * nr
161                  *  - hash of the packfile
162                  *  - file checksum
163                  * And after the 4-byte offset table might be a
164                  * variable sized table containing 8-byte entries
165                  * for offsets larger than 2^31.
166                  */
167                 size_t min_size = st_add(8 + 4*256 + hashsz + hashsz, st_mult(nr, hashsz + 4 + 4));
168                 size_t max_size = min_size;
169                 if (nr)
170                         max_size = st_add(max_size, st_mult(nr - 1, 8));
171                 if (idx_size < min_size || idx_size > max_size)
172                         return error("wrong index v2 file size in %s", path);
173                 if (idx_size != min_size &&
174                     /*
175                      * make sure we can deal with large pack offsets.
176                      * 31-bit signed offset won't be enough, neither
177                      * 32-bit unsigned one will be.
178                      */
179                     (sizeof(off_t) <= 4))
180                         return error("pack too large for current definition of off_t in %s", path);
181                 p->crc_offset = 8 + 4 * 256 + nr * hashsz;
182         }
183
184         p->index_version = version;
185         p->index_data = idx_map;
186         p->index_size = idx_size;
187         p->num_objects = nr;
188         return 0;
189 }
190
191 int open_pack_index(struct packed_git *p)
192 {
193         char *idx_name;
194         size_t len;
195         int ret;
196
197         if (p->index_data)
198                 return 0;
199
200         if (!strip_suffix(p->pack_name, ".pack", &len))
201                 BUG("pack_name does not end in .pack");
202         idx_name = xstrfmt("%.*s.idx", (int)len, p->pack_name);
203         ret = check_packed_git_idx(idx_name, p);
204         free(idx_name);
205         return ret;
206 }
207
208 uint32_t get_pack_fanout(struct packed_git *p, uint32_t value)
209 {
210         const uint32_t *level1_ofs = p->index_data;
211
212         if (!level1_ofs) {
213                 if (open_pack_index(p))
214                         return 0;
215                 level1_ofs = p->index_data;
216         }
217
218         if (p->index_version > 1) {
219                 level1_ofs += 2;
220         }
221
222         return ntohl(level1_ofs[value]);
223 }
224
225 static struct packed_git *alloc_packed_git(int extra)
226 {
227         struct packed_git *p = xmalloc(st_add(sizeof(*p), extra));
228         memset(p, 0, sizeof(*p));
229         p->pack_fd = -1;
230         return p;
231 }
232
233 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
234 {
235         const char *path = sha1_pack_name(sha1);
236         size_t alloc = st_add(strlen(path), 1);
237         struct packed_git *p = alloc_packed_git(alloc);
238
239         memcpy(p->pack_name, path, alloc); /* includes NUL */
240         hashcpy(p->hash, sha1);
241         if (check_packed_git_idx(idx_path, p)) {
242                 free(p);
243                 return NULL;
244         }
245
246         return p;
247 }
248
249 static void scan_windows(struct packed_git *p,
250         struct packed_git **lru_p,
251         struct pack_window **lru_w,
252         struct pack_window **lru_l)
253 {
254         struct pack_window *w, *w_l;
255
256         for (w_l = NULL, w = p->windows; w; w = w->next) {
257                 if (!w->inuse_cnt) {
258                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
259                                 *lru_p = p;
260                                 *lru_w = w;
261                                 *lru_l = w_l;
262                         }
263                 }
264                 w_l = w;
265         }
266 }
267
268 static int unuse_one_window(struct packed_git *current)
269 {
270         struct packed_git *p, *lru_p = NULL;
271         struct pack_window *lru_w = NULL, *lru_l = NULL;
272
273         if (current)
274                 scan_windows(current, &lru_p, &lru_w, &lru_l);
275         for (p = the_repository->objects->packed_git; p; p = p->next)
276                 scan_windows(p, &lru_p, &lru_w, &lru_l);
277         if (lru_p) {
278                 munmap(lru_w->base, lru_w->len);
279                 pack_mapped -= lru_w->len;
280                 if (lru_l)
281                         lru_l->next = lru_w->next;
282                 else
283                         lru_p->windows = lru_w->next;
284                 free(lru_w);
285                 pack_open_windows--;
286                 return 1;
287         }
288         return 0;
289 }
290
291 void close_pack_windows(struct packed_git *p)
292 {
293         while (p->windows) {
294                 struct pack_window *w = p->windows;
295
296                 if (w->inuse_cnt)
297                         die("pack '%s' still has open windows to it",
298                             p->pack_name);
299                 munmap(w->base, w->len);
300                 pack_mapped -= w->len;
301                 pack_open_windows--;
302                 p->windows = w->next;
303                 free(w);
304         }
305 }
306
307 int close_pack_fd(struct packed_git *p)
308 {
309         if (p->pack_fd < 0)
310                 return 0;
311
312         close(p->pack_fd);
313         pack_open_fds--;
314         p->pack_fd = -1;
315
316         return 1;
317 }
318
319 void close_pack_index(struct packed_git *p)
320 {
321         if (p->index_data) {
322                 munmap((void *)p->index_data, p->index_size);
323                 p->index_data = NULL;
324         }
325 }
326
327 void close_pack_revindex(struct packed_git *p) {
328         if (!p->revindex_map)
329                 return;
330
331         munmap((void *)p->revindex_map, p->revindex_size);
332         p->revindex_map = NULL;
333         p->revindex_data = NULL;
334 }
335
336 void close_pack(struct packed_git *p)
337 {
338         close_pack_windows(p);
339         close_pack_fd(p);
340         close_pack_index(p);
341         close_pack_revindex(p);
342 }
343
344 void close_object_store(struct raw_object_store *o)
345 {
346         struct packed_git *p;
347
348         for (p = o->packed_git; p; p = p->next)
349                 if (p->do_not_close)
350                         BUG("want to close pack marked 'do-not-close'");
351                 else
352                         close_pack(p);
353
354         if (o->multi_pack_index) {
355                 close_midx(o->multi_pack_index);
356                 o->multi_pack_index = NULL;
357         }
358
359         close_commit_graph(o);
360 }
361
362 void unlink_pack_path(const char *pack_name, int force_delete)
363 {
364         static const char *exts[] = {".pack", ".idx", ".rev", ".keep", ".bitmap", ".promisor"};
365         int i;
366         struct strbuf buf = STRBUF_INIT;
367         size_t plen;
368
369         strbuf_addstr(&buf, pack_name);
370         strip_suffix_mem(buf.buf, &buf.len, ".pack");
371         plen = buf.len;
372
373         if (!force_delete) {
374                 strbuf_addstr(&buf, ".keep");
375                 if (!access(buf.buf, F_OK)) {
376                         strbuf_release(&buf);
377                         return;
378                 }
379         }
380
381         for (i = 0; i < ARRAY_SIZE(exts); i++) {
382                 strbuf_setlen(&buf, plen);
383                 strbuf_addstr(&buf, exts[i]);
384                 unlink(buf.buf);
385         }
386
387         strbuf_release(&buf);
388 }
389
390 /*
391  * The LRU pack is the one with the oldest MRU window, preferring packs
392  * with no used windows, or the oldest mtime if it has no windows allocated.
393  */
394 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
395 {
396         struct pack_window *w, *this_mru_w;
397         int has_windows_inuse = 0;
398
399         /*
400          * Reject this pack if it has windows and the previously selected
401          * one does not.  If this pack does not have windows, reject
402          * it if the pack file is newer than the previously selected one.
403          */
404         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
405                 return;
406
407         for (w = this_mru_w = p->windows; w; w = w->next) {
408                 /*
409                  * Reject this pack if any of its windows are in use,
410                  * but the previously selected pack did not have any
411                  * inuse windows.  Otherwise, record that this pack
412                  * has windows in use.
413                  */
414                 if (w->inuse_cnt) {
415                         if (*accept_windows_inuse)
416                                 has_windows_inuse = 1;
417                         else
418                                 return;
419                 }
420
421                 if (w->last_used > this_mru_w->last_used)
422                         this_mru_w = w;
423
424                 /*
425                  * Reject this pack if it has windows that have been
426                  * used more recently than the previously selected pack.
427                  * If the previously selected pack had windows inuse and
428                  * we have not encountered a window in this pack that is
429                  * inuse, skip this check since we prefer a pack with no
430                  * inuse windows to one that has inuse windows.
431                  */
432                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
433                     this_mru_w->last_used > (*mru_w)->last_used)
434                         return;
435         }
436
437         /*
438          * Select this pack.
439          */
440         *mru_w = this_mru_w;
441         *lru_p = p;
442         *accept_windows_inuse = has_windows_inuse;
443 }
444
445 static int close_one_pack(void)
446 {
447         struct packed_git *p, *lru_p = NULL;
448         struct pack_window *mru_w = NULL;
449         int accept_windows_inuse = 1;
450
451         for (p = the_repository->objects->packed_git; p; p = p->next) {
452                 if (p->pack_fd == -1)
453                         continue;
454                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
455         }
456
457         if (lru_p)
458                 return close_pack_fd(lru_p);
459
460         return 0;
461 }
462
463 static unsigned int get_max_fd_limit(void)
464 {
465 #ifdef RLIMIT_NOFILE
466         {
467                 struct rlimit lim;
468
469                 if (!getrlimit(RLIMIT_NOFILE, &lim))
470                         return lim.rlim_cur;
471         }
472 #endif
473
474 #ifdef _SC_OPEN_MAX
475         {
476                 long open_max = sysconf(_SC_OPEN_MAX);
477                 if (0 < open_max)
478                         return open_max;
479                 /*
480                  * Otherwise, we got -1 for one of the two
481                  * reasons:
482                  *
483                  * (1) sysconf() did not understand _SC_OPEN_MAX
484                  *     and signaled an error with -1; or
485                  * (2) sysconf() said there is no limit.
486                  *
487                  * We _could_ clear errno before calling sysconf() to
488                  * tell these two cases apart and return a huge number
489                  * in the latter case to let the caller cap it to a
490                  * value that is not so selfish, but letting the
491                  * fallback OPEN_MAX codepath take care of these cases
492                  * is a lot simpler.
493                  */
494         }
495 #endif
496
497 #ifdef OPEN_MAX
498         return OPEN_MAX;
499 #else
500         return 1; /* see the caller ;-) */
501 #endif
502 }
503
504 const char *pack_basename(struct packed_git *p)
505 {
506         const char *ret = strrchr(p->pack_name, '/');
507         if (ret)
508                 ret = ret + 1; /* skip past slash */
509         else
510                 ret = p->pack_name; /* we only have a base */
511         return ret;
512 }
513
514 /*
515  * Do not call this directly as this leaks p->pack_fd on error return;
516  * call open_packed_git() instead.
517  */
518 static int open_packed_git_1(struct packed_git *p)
519 {
520         struct stat st;
521         struct pack_header hdr;
522         unsigned char hash[GIT_MAX_RAWSZ];
523         unsigned char *idx_hash;
524         ssize_t read_result;
525         const unsigned hashsz = the_hash_algo->rawsz;
526
527         if (open_pack_index(p))
528                 return error("packfile %s index unavailable", p->pack_name);
529
530         if (!pack_max_fds) {
531                 unsigned int max_fds = get_max_fd_limit();
532
533                 /* Save 3 for stdin/stdout/stderr, 22 for work */
534                 if (25 < max_fds)
535                         pack_max_fds = max_fds - 25;
536                 else
537                         pack_max_fds = 1;
538         }
539
540         while (pack_max_fds <= pack_open_fds && close_one_pack())
541                 ; /* nothing */
542
543         p->pack_fd = git_open(p->pack_name);
544         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
545                 return -1;
546         pack_open_fds++;
547
548         /* If we created the struct before we had the pack we lack size. */
549         if (!p->pack_size) {
550                 if (!S_ISREG(st.st_mode))
551                         return error("packfile %s not a regular file", p->pack_name);
552                 p->pack_size = st.st_size;
553         } else if (p->pack_size != st.st_size)
554                 return error("packfile %s size changed", p->pack_name);
555
556         /* Verify we recognize this pack file format. */
557         read_result = read_in_full(p->pack_fd, &hdr, sizeof(hdr));
558         if (read_result < 0)
559                 return error_errno("error reading from %s", p->pack_name);
560         if (read_result != sizeof(hdr))
561                 return error("file %s is far too short to be a packfile", p->pack_name);
562         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
563                 return error("file %s is not a GIT packfile", p->pack_name);
564         if (!pack_version_ok(hdr.hdr_version))
565                 return error("packfile %s is version %"PRIu32" and not"
566                         " supported (try upgrading GIT to a newer version)",
567                         p->pack_name, ntohl(hdr.hdr_version));
568
569         /* Verify the pack matches its index. */
570         if (p->num_objects != ntohl(hdr.hdr_entries))
571                 return error("packfile %s claims to have %"PRIu32" objects"
572                              " while index indicates %"PRIu32" objects",
573                              p->pack_name, ntohl(hdr.hdr_entries),
574                              p->num_objects);
575         read_result = pread_in_full(p->pack_fd, hash, hashsz,
576                                         p->pack_size - hashsz);
577         if (read_result < 0)
578                 return error_errno("error reading from %s", p->pack_name);
579         if (read_result != hashsz)
580                 return error("packfile %s signature is unavailable", p->pack_name);
581         idx_hash = ((unsigned char *)p->index_data) + p->index_size - hashsz * 2;
582         if (!hasheq(hash, idx_hash))
583                 return error("packfile %s does not match index", p->pack_name);
584         return 0;
585 }
586
587 static int open_packed_git(struct packed_git *p)
588 {
589         if (!open_packed_git_1(p))
590                 return 0;
591         close_pack_fd(p);
592         return -1;
593 }
594
595 static int in_window(struct pack_window *win, off_t offset)
596 {
597         /* We must promise at least one full hash after the
598          * offset is available from this window, otherwise the offset
599          * is not actually in this window and a different window (which
600          * has that one hash excess) must be used.  This is to support
601          * the object header and delta base parsing routines below.
602          */
603         off_t win_off = win->offset;
604         return win_off <= offset
605                 && (offset + the_hash_algo->rawsz) <= (win_off + win->len);
606 }
607
608 unsigned char *use_pack(struct packed_git *p,
609                 struct pack_window **w_cursor,
610                 off_t offset,
611                 unsigned long *left)
612 {
613         struct pack_window *win = *w_cursor;
614
615         /* Since packfiles end in a hash of their content and it's
616          * pointless to ask for an offset into the middle of that
617          * hash, and the in_window function above wouldn't match
618          * don't allow an offset too close to the end of the file.
619          */
620         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
621                 die("packfile %s cannot be accessed", p->pack_name);
622         if (offset > (p->pack_size - the_hash_algo->rawsz))
623                 die("offset beyond end of packfile (truncated pack?)");
624         if (offset < 0)
625                 die(_("offset before end of packfile (broken .idx?)"));
626
627         if (!win || !in_window(win, offset)) {
628                 if (win)
629                         win->inuse_cnt--;
630                 for (win = p->windows; win; win = win->next) {
631                         if (in_window(win, offset))
632                                 break;
633                 }
634                 if (!win) {
635                         size_t window_align = packed_git_window_size / 2;
636                         off_t len;
637
638                         if (p->pack_fd == -1 && open_packed_git(p))
639                                 die("packfile %s cannot be accessed", p->pack_name);
640
641                         win = xcalloc(1, sizeof(*win));
642                         win->offset = (offset / window_align) * window_align;
643                         len = p->pack_size - win->offset;
644                         if (len > packed_git_window_size)
645                                 len = packed_git_window_size;
646                         win->len = (size_t)len;
647                         pack_mapped += win->len;
648                         while (packed_git_limit < pack_mapped
649                                 && unuse_one_window(p))
650                                 ; /* nothing */
651                         win->base = xmmap_gently(NULL, win->len,
652                                 PROT_READ, MAP_PRIVATE,
653                                 p->pack_fd, win->offset);
654                         if (win->base == MAP_FAILED)
655                                 die_errno("packfile %s cannot be mapped",
656                                           p->pack_name);
657                         if (!win->offset && win->len == p->pack_size
658                                 && !p->do_not_close)
659                                 close_pack_fd(p);
660                         pack_mmap_calls++;
661                         pack_open_windows++;
662                         if (pack_mapped > peak_pack_mapped)
663                                 peak_pack_mapped = pack_mapped;
664                         if (pack_open_windows > peak_pack_open_windows)
665                                 peak_pack_open_windows = pack_open_windows;
666                         win->next = p->windows;
667                         p->windows = win;
668                 }
669         }
670         if (win != *w_cursor) {
671                 win->last_used = pack_used_ctr++;
672                 win->inuse_cnt++;
673                 *w_cursor = win;
674         }
675         offset -= win->offset;
676         if (left)
677                 *left = win->len - xsize_t(offset);
678         return win->base + offset;
679 }
680
681 void unuse_pack(struct pack_window **w_cursor)
682 {
683         struct pack_window *w = *w_cursor;
684         if (w) {
685                 w->inuse_cnt--;
686                 *w_cursor = NULL;
687         }
688 }
689
690 struct packed_git *add_packed_git(const char *path, size_t path_len, int local)
691 {
692         struct stat st;
693         size_t alloc;
694         struct packed_git *p;
695
696         /*
697          * Make sure a corresponding .pack file exists and that
698          * the index looks sane.
699          */
700         if (!strip_suffix_mem(path, &path_len, ".idx"))
701                 return NULL;
702
703         /*
704          * ".promisor" is long enough to hold any suffix we're adding (and
705          * the use xsnprintf double-checks that)
706          */
707         alloc = st_add3(path_len, strlen(".promisor"), 1);
708         p = alloc_packed_git(alloc);
709         memcpy(p->pack_name, path, path_len);
710
711         xsnprintf(p->pack_name + path_len, alloc - path_len, ".keep");
712         if (!access(p->pack_name, F_OK))
713                 p->pack_keep = 1;
714
715         xsnprintf(p->pack_name + path_len, alloc - path_len, ".promisor");
716         if (!access(p->pack_name, F_OK))
717                 p->pack_promisor = 1;
718
719         xsnprintf(p->pack_name + path_len, alloc - path_len, ".pack");
720         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
721                 free(p);
722                 return NULL;
723         }
724
725         /* ok, it looks sane as far as we can check without
726          * actually mapping the pack file.
727          */
728         p->pack_size = st.st_size;
729         p->pack_local = local;
730         p->mtime = st.st_mtime;
731         if (path_len < the_hash_algo->hexsz ||
732             get_sha1_hex(path + path_len - the_hash_algo->hexsz, p->hash))
733                 hashclr(p->hash);
734         return p;
735 }
736
737 void install_packed_git(struct repository *r, struct packed_git *pack)
738 {
739         if (pack->pack_fd != -1)
740                 pack_open_fds++;
741
742         pack->next = r->objects->packed_git;
743         r->objects->packed_git = pack;
744
745         hashmap_entry_init(&pack->packmap_ent, strhash(pack->pack_name));
746         hashmap_add(&r->objects->pack_map, &pack->packmap_ent);
747 }
748
749 void (*report_garbage)(unsigned seen_bits, const char *path);
750
751 static void report_helper(const struct string_list *list,
752                           int seen_bits, int first, int last)
753 {
754         if (seen_bits == (PACKDIR_FILE_PACK|PACKDIR_FILE_IDX))
755                 return;
756
757         for (; first < last; first++)
758                 report_garbage(seen_bits, list->items[first].string);
759 }
760
761 static void report_pack_garbage(struct string_list *list)
762 {
763         int i, baselen = -1, first = 0, seen_bits = 0;
764
765         if (!report_garbage)
766                 return;
767
768         string_list_sort(list);
769
770         for (i = 0; i < list->nr; i++) {
771                 const char *path = list->items[i].string;
772                 if (baselen != -1 &&
773                     strncmp(path, list->items[first].string, baselen)) {
774                         report_helper(list, seen_bits, first, i);
775                         baselen = -1;
776                         seen_bits = 0;
777                 }
778                 if (baselen == -1) {
779                         const char *dot = strrchr(path, '.');
780                         if (!dot) {
781                                 report_garbage(PACKDIR_FILE_GARBAGE, path);
782                                 continue;
783                         }
784                         baselen = dot - path + 1;
785                         first = i;
786                 }
787                 if (!strcmp(path + baselen, "pack"))
788                         seen_bits |= 1;
789                 else if (!strcmp(path + baselen, "idx"))
790                         seen_bits |= 2;
791         }
792         report_helper(list, seen_bits, first, list->nr);
793 }
794
795 void for_each_file_in_pack_dir(const char *objdir,
796                                each_file_in_pack_dir_fn fn,
797                                void *data)
798 {
799         struct strbuf path = STRBUF_INIT;
800         size_t dirnamelen;
801         DIR *dir;
802         struct dirent *de;
803
804         strbuf_addstr(&path, objdir);
805         strbuf_addstr(&path, "/pack");
806         dir = opendir(path.buf);
807         if (!dir) {
808                 if (errno != ENOENT)
809                         error_errno("unable to open object pack directory: %s",
810                                     path.buf);
811                 strbuf_release(&path);
812                 return;
813         }
814         strbuf_addch(&path, '/');
815         dirnamelen = path.len;
816         while ((de = readdir(dir)) != NULL) {
817                 if (is_dot_or_dotdot(de->d_name))
818                         continue;
819
820                 strbuf_setlen(&path, dirnamelen);
821                 strbuf_addstr(&path, de->d_name);
822
823                 fn(path.buf, path.len, de->d_name, data);
824         }
825
826         closedir(dir);
827         strbuf_release(&path);
828 }
829
830 struct prepare_pack_data {
831         struct repository *r;
832         struct string_list *garbage;
833         int local;
834         struct multi_pack_index *m;
835 };
836
837 static void prepare_pack(const char *full_name, size_t full_name_len,
838                          const char *file_name, void *_data)
839 {
840         struct prepare_pack_data *data = (struct prepare_pack_data *)_data;
841         struct packed_git *p;
842         size_t base_len = full_name_len;
843
844         if (strip_suffix_mem(full_name, &base_len, ".idx") &&
845             !(data->m && midx_contains_pack(data->m, file_name))) {
846                 struct hashmap_entry hent;
847                 char *pack_name = xstrfmt("%.*s.pack", (int)base_len, full_name);
848                 unsigned int hash = strhash(pack_name);
849                 hashmap_entry_init(&hent, hash);
850
851                 /* Don't reopen a pack we already have. */
852                 if (!hashmap_get(&data->r->objects->pack_map, &hent, pack_name)) {
853                         p = add_packed_git(full_name, full_name_len, data->local);
854                         if (p)
855                                 install_packed_git(data->r, p);
856                 }
857                 free(pack_name);
858         }
859
860         if (!report_garbage)
861                 return;
862
863         if (!strcmp(file_name, "multi-pack-index"))
864                 return;
865         if (ends_with(file_name, ".idx") ||
866             ends_with(file_name, ".rev") ||
867             ends_with(file_name, ".pack") ||
868             ends_with(file_name, ".bitmap") ||
869             ends_with(file_name, ".keep") ||
870             ends_with(file_name, ".promisor"))
871                 string_list_append(data->garbage, full_name);
872         else
873                 report_garbage(PACKDIR_FILE_GARBAGE, full_name);
874 }
875
876 static void prepare_packed_git_one(struct repository *r, char *objdir, int local)
877 {
878         struct prepare_pack_data data;
879         struct string_list garbage = STRING_LIST_INIT_DUP;
880
881         data.m = r->objects->multi_pack_index;
882
883         /* look for the multi-pack-index for this object directory */
884         while (data.m && strcmp(data.m->object_dir, objdir))
885                 data.m = data.m->next;
886
887         data.r = r;
888         data.garbage = &garbage;
889         data.local = local;
890
891         for_each_file_in_pack_dir(objdir, prepare_pack, &data);
892
893         report_pack_garbage(data.garbage);
894         string_list_clear(data.garbage, 0);
895 }
896
897 static void prepare_packed_git(struct repository *r);
898 /*
899  * Give a fast, rough count of the number of objects in the repository. This
900  * ignores loose objects completely. If you have a lot of them, then either
901  * you should repack because your performance will be awful, or they are
902  * all unreachable objects about to be pruned, in which case they're not really
903  * interesting as a measure of repo size in the first place.
904  */
905 unsigned long repo_approximate_object_count(struct repository *r)
906 {
907         if (!r->objects->approximate_object_count_valid) {
908                 unsigned long count;
909                 struct multi_pack_index *m;
910                 struct packed_git *p;
911
912                 prepare_packed_git(r);
913                 count = 0;
914                 for (m = get_multi_pack_index(r); m; m = m->next)
915                         count += m->num_objects;
916                 for (p = r->objects->packed_git; p; p = p->next) {
917                         if (open_pack_index(p))
918                                 continue;
919                         count += p->num_objects;
920                 }
921                 r->objects->approximate_object_count = count;
922                 r->objects->approximate_object_count_valid = 1;
923         }
924         return r->objects->approximate_object_count;
925 }
926
927 static void *get_next_packed_git(const void *p)
928 {
929         return ((const struct packed_git *)p)->next;
930 }
931
932 static void set_next_packed_git(void *p, void *next)
933 {
934         ((struct packed_git *)p)->next = next;
935 }
936
937 static int sort_pack(const void *a_, const void *b_)
938 {
939         const struct packed_git *a = a_;
940         const struct packed_git *b = b_;
941         int st;
942
943         /*
944          * Local packs tend to contain objects specific to our
945          * variant of the project than remote ones.  In addition,
946          * remote ones could be on a network mounted filesystem.
947          * Favor local ones for these reasons.
948          */
949         st = a->pack_local - b->pack_local;
950         if (st)
951                 return -st;
952
953         /*
954          * Younger packs tend to contain more recent objects,
955          * and more recent objects tend to get accessed more
956          * often.
957          */
958         if (a->mtime < b->mtime)
959                 return 1;
960         else if (a->mtime == b->mtime)
961                 return 0;
962         return -1;
963 }
964
965 static void rearrange_packed_git(struct repository *r)
966 {
967         r->objects->packed_git = llist_mergesort(
968                 r->objects->packed_git, get_next_packed_git,
969                 set_next_packed_git, sort_pack);
970 }
971
972 static void prepare_packed_git_mru(struct repository *r)
973 {
974         struct packed_git *p;
975
976         INIT_LIST_HEAD(&r->objects->packed_git_mru);
977
978         for (p = r->objects->packed_git; p; p = p->next)
979                 list_add_tail(&p->mru, &r->objects->packed_git_mru);
980 }
981
982 static void prepare_packed_git(struct repository *r)
983 {
984         struct object_directory *odb;
985
986         if (r->objects->packed_git_initialized)
987                 return;
988
989         prepare_alt_odb(r);
990         for (odb = r->objects->odb; odb; odb = odb->next) {
991                 int local = (odb == r->objects->odb);
992                 prepare_multi_pack_index_one(r, odb->path, local);
993                 prepare_packed_git_one(r, odb->path, local);
994         }
995         rearrange_packed_git(r);
996
997         prepare_packed_git_mru(r);
998         r->objects->packed_git_initialized = 1;
999 }
1000
1001 void reprepare_packed_git(struct repository *r)
1002 {
1003         struct object_directory *odb;
1004
1005         obj_read_lock();
1006         for (odb = r->objects->odb; odb; odb = odb->next)
1007                 odb_clear_loose_cache(odb);
1008
1009         r->objects->approximate_object_count_valid = 0;
1010         r->objects->packed_git_initialized = 0;
1011         prepare_packed_git(r);
1012         obj_read_unlock();
1013 }
1014
1015 struct packed_git *get_packed_git(struct repository *r)
1016 {
1017         prepare_packed_git(r);
1018         return r->objects->packed_git;
1019 }
1020
1021 struct multi_pack_index *get_multi_pack_index(struct repository *r)
1022 {
1023         prepare_packed_git(r);
1024         return r->objects->multi_pack_index;
1025 }
1026
1027 struct multi_pack_index *get_local_multi_pack_index(struct repository *r)
1028 {
1029         struct multi_pack_index *m = get_multi_pack_index(r);
1030
1031         /* no need to iterate; we always put the local one first (if any) */
1032         if (m && m->local)
1033                 return m;
1034
1035         return NULL;
1036 }
1037
1038 struct packed_git *get_all_packs(struct repository *r)
1039 {
1040         struct multi_pack_index *m;
1041
1042         prepare_packed_git(r);
1043         for (m = r->objects->multi_pack_index; m; m = m->next) {
1044                 uint32_t i;
1045                 for (i = 0; i < m->num_packs; i++)
1046                         prepare_midx_pack(r, m, i);
1047         }
1048
1049         return r->objects->packed_git;
1050 }
1051
1052 struct list_head *get_packed_git_mru(struct repository *r)
1053 {
1054         prepare_packed_git(r);
1055         return &r->objects->packed_git_mru;
1056 }
1057
1058 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1059                 unsigned long len, enum object_type *type, unsigned long *sizep)
1060 {
1061         unsigned shift;
1062         unsigned long size, c;
1063         unsigned long used = 0;
1064
1065         c = buf[used++];
1066         *type = (c >> 4) & 7;
1067         size = c & 15;
1068         shift = 4;
1069         while (c & 0x80) {
1070                 if (len <= used || bitsizeof(long) <= shift) {
1071                         error("bad object header");
1072                         size = used = 0;
1073                         break;
1074                 }
1075                 c = buf[used++];
1076                 size += (c & 0x7f) << shift;
1077                 shift += 7;
1078         }
1079         *sizep = size;
1080         return used;
1081 }
1082
1083 unsigned long get_size_from_delta(struct packed_git *p,
1084                                   struct pack_window **w_curs,
1085                                   off_t curpos)
1086 {
1087         const unsigned char *data;
1088         unsigned char delta_head[20], *in;
1089         git_zstream stream;
1090         int st;
1091
1092         memset(&stream, 0, sizeof(stream));
1093         stream.next_out = delta_head;
1094         stream.avail_out = sizeof(delta_head);
1095
1096         git_inflate_init(&stream);
1097         do {
1098                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1099                 stream.next_in = in;
1100                 /*
1101                  * Note: the window section returned by use_pack() must be
1102                  * available throughout git_inflate()'s unlocked execution. To
1103                  * ensure no other thread will modify the window in the
1104                  * meantime, we rely on the packed_window.inuse_cnt. This
1105                  * counter is incremented before window reading and checked
1106                  * before window disposal.
1107                  *
1108                  * Other worrying sections could be the call to close_pack_fd(),
1109                  * which can close packs even with in-use windows, and to
1110                  * reprepare_packed_git(). Regarding the former, mmap doc says:
1111                  * "closing the file descriptor does not unmap the region". And
1112                  * for the latter, it won't re-open already available packs.
1113                  */
1114                 obj_read_unlock();
1115                 st = git_inflate(&stream, Z_FINISH);
1116                 obj_read_lock();
1117                 curpos += stream.next_in - in;
1118         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1119                  stream.total_out < sizeof(delta_head));
1120         git_inflate_end(&stream);
1121         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1122                 error("delta data unpack-initial failed");
1123                 return 0;
1124         }
1125
1126         /* Examine the initial part of the delta to figure out
1127          * the result size.
1128          */
1129         data = delta_head;
1130
1131         /* ignore base size */
1132         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1133
1134         /* Read the result size */
1135         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1136 }
1137
1138 int unpack_object_header(struct packed_git *p,
1139                          struct pack_window **w_curs,
1140                          off_t *curpos,
1141                          unsigned long *sizep)
1142 {
1143         unsigned char *base;
1144         unsigned long left;
1145         unsigned long used;
1146         enum object_type type;
1147
1148         /* use_pack() assures us we have [base, base + 20) available
1149          * as a range that we can look at.  (Its actually the hash
1150          * size that is assured.)  With our object header encoding
1151          * the maximum deflated object size is 2^137, which is just
1152          * insane, so we know won't exceed what we have been given.
1153          */
1154         base = use_pack(p, w_curs, *curpos, &left);
1155         used = unpack_object_header_buffer(base, left, &type, sizep);
1156         if (!used) {
1157                 type = OBJ_BAD;
1158         } else
1159                 *curpos += used;
1160
1161         return type;
1162 }
1163
1164 void mark_bad_packed_object(struct packed_git *p, const unsigned char *sha1)
1165 {
1166         unsigned i;
1167         const unsigned hashsz = the_hash_algo->rawsz;
1168         for (i = 0; i < p->num_bad_objects; i++)
1169                 if (hasheq(sha1, p->bad_object_sha1 + hashsz * i))
1170                         return;
1171         p->bad_object_sha1 = xrealloc(p->bad_object_sha1,
1172                                       st_mult(GIT_MAX_RAWSZ,
1173                                               st_add(p->num_bad_objects, 1)));
1174         hashcpy(p->bad_object_sha1 + hashsz * p->num_bad_objects, sha1);
1175         p->num_bad_objects++;
1176 }
1177
1178 const struct packed_git *has_packed_and_bad(struct repository *r,
1179                                             const unsigned char *sha1)
1180 {
1181         struct packed_git *p;
1182         unsigned i;
1183
1184         for (p = r->objects->packed_git; p; p = p->next)
1185                 for (i = 0; i < p->num_bad_objects; i++)
1186                         if (hasheq(sha1,
1187                                    p->bad_object_sha1 + the_hash_algo->rawsz * i))
1188                                 return p;
1189         return NULL;
1190 }
1191
1192 off_t get_delta_base(struct packed_git *p,
1193                      struct pack_window **w_curs,
1194                      off_t *curpos,
1195                      enum object_type type,
1196                      off_t delta_obj_offset)
1197 {
1198         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1199         off_t base_offset;
1200
1201         /* use_pack() assured us we have [base_info, base_info + 20)
1202          * as a range that we can look at without walking off the
1203          * end of the mapped window.  Its actually the hash size
1204          * that is assured.  An OFS_DELTA longer than the hash size
1205          * is stupid, as then a REF_DELTA would be smaller to store.
1206          */
1207         if (type == OBJ_OFS_DELTA) {
1208                 unsigned used = 0;
1209                 unsigned char c = base_info[used++];
1210                 base_offset = c & 127;
1211                 while (c & 128) {
1212                         base_offset += 1;
1213                         if (!base_offset || MSB(base_offset, 7))
1214                                 return 0;  /* overflow */
1215                         c = base_info[used++];
1216                         base_offset = (base_offset << 7) + (c & 127);
1217                 }
1218                 base_offset = delta_obj_offset - base_offset;
1219                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1220                         return 0;  /* out of bound */
1221                 *curpos += used;
1222         } else if (type == OBJ_REF_DELTA) {
1223                 /* The base entry _must_ be in the same pack */
1224                 base_offset = find_pack_entry_one(base_info, p);
1225                 *curpos += the_hash_algo->rawsz;
1226         } else
1227                 die("I am totally screwed");
1228         return base_offset;
1229 }
1230
1231 /*
1232  * Like get_delta_base above, but we return the sha1 instead of the pack
1233  * offset. This means it is cheaper for REF deltas (we do not have to do
1234  * the final object lookup), but more expensive for OFS deltas (we
1235  * have to load the revidx to convert the offset back into a sha1).
1236  */
1237 static int get_delta_base_oid(struct packed_git *p,
1238                               struct pack_window **w_curs,
1239                               off_t curpos,
1240                               struct object_id *oid,
1241                               enum object_type type,
1242                               off_t delta_obj_offset)
1243 {
1244         if (type == OBJ_REF_DELTA) {
1245                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
1246                 oidread(oid, base);
1247                 return 0;
1248         } else if (type == OBJ_OFS_DELTA) {
1249                 uint32_t base_pos;
1250                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
1251                                                    type, delta_obj_offset);
1252
1253                 if (!base_offset)
1254                         return -1;
1255
1256                 if (offset_to_pack_pos(p, base_offset, &base_pos) < 0)
1257                         return -1;
1258
1259                 return nth_packed_object_id(oid, p,
1260                                             pack_pos_to_index(p, base_pos));
1261         } else
1262                 return -1;
1263 }
1264
1265 static int retry_bad_packed_offset(struct repository *r,
1266                                    struct packed_git *p,
1267                                    off_t obj_offset)
1268 {
1269         int type;
1270         uint32_t pos;
1271         struct object_id oid;
1272         if (offset_to_pack_pos(p, obj_offset, &pos) < 0)
1273                 return OBJ_BAD;
1274         nth_packed_object_id(&oid, p, pack_pos_to_index(p, pos));
1275         mark_bad_packed_object(p, oid.hash);
1276         type = oid_object_info(r, &oid, NULL);
1277         if (type <= OBJ_NONE)
1278                 return OBJ_BAD;
1279         return type;
1280 }
1281
1282 #define POI_STACK_PREALLOC 64
1283
1284 static enum object_type packed_to_object_type(struct repository *r,
1285                                               struct packed_git *p,
1286                                               off_t obj_offset,
1287                                               enum object_type type,
1288                                               struct pack_window **w_curs,
1289                                               off_t curpos)
1290 {
1291         off_t small_poi_stack[POI_STACK_PREALLOC];
1292         off_t *poi_stack = small_poi_stack;
1293         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
1294
1295         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1296                 off_t base_offset;
1297                 unsigned long size;
1298                 /* Push the object we're going to leave behind */
1299                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
1300                         poi_stack_alloc = alloc_nr(poi_stack_nr);
1301                         ALLOC_ARRAY(poi_stack, poi_stack_alloc);
1302                         COPY_ARRAY(poi_stack, small_poi_stack, poi_stack_nr);
1303                 } else {
1304                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
1305                 }
1306                 poi_stack[poi_stack_nr++] = obj_offset;
1307                 /* If parsing the base offset fails, just unwind */
1308                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1309                 if (!base_offset)
1310                         goto unwind;
1311                 curpos = obj_offset = base_offset;
1312                 type = unpack_object_header(p, w_curs, &curpos, &size);
1313                 if (type <= OBJ_NONE) {
1314                         /* If getting the base itself fails, we first
1315                          * retry the base, otherwise unwind */
1316                         type = retry_bad_packed_offset(r, p, base_offset);
1317                         if (type > OBJ_NONE)
1318                                 goto out;
1319                         goto unwind;
1320                 }
1321         }
1322
1323         switch (type) {
1324         case OBJ_BAD:
1325         case OBJ_COMMIT:
1326         case OBJ_TREE:
1327         case OBJ_BLOB:
1328         case OBJ_TAG:
1329                 break;
1330         default:
1331                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1332                       type, (uintmax_t)obj_offset, p->pack_name);
1333                 type = OBJ_BAD;
1334         }
1335
1336 out:
1337         if (poi_stack != small_poi_stack)
1338                 free(poi_stack);
1339         return type;
1340
1341 unwind:
1342         while (poi_stack_nr) {
1343                 obj_offset = poi_stack[--poi_stack_nr];
1344                 type = retry_bad_packed_offset(r, p, obj_offset);
1345                 if (type > OBJ_NONE)
1346                         goto out;
1347         }
1348         type = OBJ_BAD;
1349         goto out;
1350 }
1351
1352 static struct hashmap delta_base_cache;
1353 static size_t delta_base_cached;
1354
1355 static LIST_HEAD(delta_base_cache_lru);
1356
1357 struct delta_base_cache_key {
1358         struct packed_git *p;
1359         off_t base_offset;
1360 };
1361
1362 struct delta_base_cache_entry {
1363         struct hashmap_entry ent;
1364         struct delta_base_cache_key key;
1365         struct list_head lru;
1366         void *data;
1367         unsigned long size;
1368         enum object_type type;
1369 };
1370
1371 static unsigned int pack_entry_hash(struct packed_git *p, off_t base_offset)
1372 {
1373         unsigned int hash;
1374
1375         hash = (unsigned int)(intptr_t)p + (unsigned int)base_offset;
1376         hash += (hash >> 8) + (hash >> 16);
1377         return hash;
1378 }
1379
1380 static struct delta_base_cache_entry *
1381 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
1382 {
1383         struct hashmap_entry entry, *e;
1384         struct delta_base_cache_key key;
1385
1386         if (!delta_base_cache.cmpfn)
1387                 return NULL;
1388
1389         hashmap_entry_init(&entry, pack_entry_hash(p, base_offset));
1390         key.p = p;
1391         key.base_offset = base_offset;
1392         e = hashmap_get(&delta_base_cache, &entry, &key);
1393         return e ? container_of(e, struct delta_base_cache_entry, ent) : NULL;
1394 }
1395
1396 static int delta_base_cache_key_eq(const struct delta_base_cache_key *a,
1397                                    const struct delta_base_cache_key *b)
1398 {
1399         return a->p == b->p && a->base_offset == b->base_offset;
1400 }
1401
1402 static int delta_base_cache_hash_cmp(const void *unused_cmp_data,
1403                                      const struct hashmap_entry *va,
1404                                      const struct hashmap_entry *vb,
1405                                      const void *vkey)
1406 {
1407         const struct delta_base_cache_entry *a, *b;
1408         const struct delta_base_cache_key *key = vkey;
1409
1410         a = container_of(va, const struct delta_base_cache_entry, ent);
1411         b = container_of(vb, const struct delta_base_cache_entry, ent);
1412
1413         if (key)
1414                 return !delta_base_cache_key_eq(&a->key, key);
1415         else
1416                 return !delta_base_cache_key_eq(&a->key, &b->key);
1417 }
1418
1419 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
1420 {
1421         return !!get_delta_base_cache_entry(p, base_offset);
1422 }
1423
1424 /*
1425  * Remove the entry from the cache, but do _not_ free the associated
1426  * entry data. The caller takes ownership of the "data" buffer, and
1427  * should copy out any fields it wants before detaching.
1428  */
1429 static void detach_delta_base_cache_entry(struct delta_base_cache_entry *ent)
1430 {
1431         hashmap_remove(&delta_base_cache, &ent->ent, &ent->key);
1432         list_del(&ent->lru);
1433         delta_base_cached -= ent->size;
1434         free(ent);
1435 }
1436
1437 static void *cache_or_unpack_entry(struct repository *r, struct packed_git *p,
1438                                    off_t base_offset, unsigned long *base_size,
1439                                    enum object_type *type)
1440 {
1441         struct delta_base_cache_entry *ent;
1442
1443         ent = get_delta_base_cache_entry(p, base_offset);
1444         if (!ent)
1445                 return unpack_entry(r, p, base_offset, type, base_size);
1446
1447         if (type)
1448                 *type = ent->type;
1449         if (base_size)
1450                 *base_size = ent->size;
1451         return xmemdupz(ent->data, ent->size);
1452 }
1453
1454 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
1455 {
1456         free(ent->data);
1457         detach_delta_base_cache_entry(ent);
1458 }
1459
1460 void clear_delta_base_cache(void)
1461 {
1462         struct list_head *lru, *tmp;
1463         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
1464                 struct delta_base_cache_entry *entry =
1465                         list_entry(lru, struct delta_base_cache_entry, lru);
1466                 release_delta_base_cache(entry);
1467         }
1468 }
1469
1470 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
1471         void *base, unsigned long base_size, enum object_type type)
1472 {
1473         struct delta_base_cache_entry *ent;
1474         struct list_head *lru, *tmp;
1475
1476         /*
1477          * Check required to avoid redundant entries when more than one thread
1478          * is unpacking the same object, in unpack_entry() (since its phases I
1479          * and III might run concurrently across multiple threads).
1480          */
1481         if (in_delta_base_cache(p, base_offset)) {
1482                 free(base);
1483                 return;
1484         }
1485
1486         delta_base_cached += base_size;
1487
1488         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
1489                 struct delta_base_cache_entry *f =
1490                         list_entry(lru, struct delta_base_cache_entry, lru);
1491                 if (delta_base_cached <= delta_base_cache_limit)
1492                         break;
1493                 release_delta_base_cache(f);
1494         }
1495
1496         ent = xmalloc(sizeof(*ent));
1497         ent->key.p = p;
1498         ent->key.base_offset = base_offset;
1499         ent->type = type;
1500         ent->data = base;
1501         ent->size = base_size;
1502         list_add_tail(&ent->lru, &delta_base_cache_lru);
1503
1504         if (!delta_base_cache.cmpfn)
1505                 hashmap_init(&delta_base_cache, delta_base_cache_hash_cmp, NULL, 0);
1506         hashmap_entry_init(&ent->ent, pack_entry_hash(p, base_offset));
1507         hashmap_add(&delta_base_cache, &ent->ent);
1508 }
1509
1510 int packed_object_info(struct repository *r, struct packed_git *p,
1511                        off_t obj_offset, struct object_info *oi)
1512 {
1513         struct pack_window *w_curs = NULL;
1514         unsigned long size;
1515         off_t curpos = obj_offset;
1516         enum object_type type;
1517
1518         /*
1519          * We always get the representation type, but only convert it to
1520          * a "real" type later if the caller is interested.
1521          */
1522         if (oi->contentp) {
1523                 *oi->contentp = cache_or_unpack_entry(r, p, obj_offset, oi->sizep,
1524                                                       &type);
1525                 if (!*oi->contentp)
1526                         type = OBJ_BAD;
1527         } else {
1528                 type = unpack_object_header(p, &w_curs, &curpos, &size);
1529         }
1530
1531         if (!oi->contentp && oi->sizep) {
1532                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1533                         off_t tmp_pos = curpos;
1534                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
1535                                                            type, obj_offset);
1536                         if (!base_offset) {
1537                                 type = OBJ_BAD;
1538                                 goto out;
1539                         }
1540                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
1541                         if (*oi->sizep == 0) {
1542                                 type = OBJ_BAD;
1543                                 goto out;
1544                         }
1545                 } else {
1546                         *oi->sizep = size;
1547                 }
1548         }
1549
1550         if (oi->disk_sizep) {
1551                 uint32_t pos;
1552                 if (offset_to_pack_pos(p, obj_offset, &pos) < 0) {
1553                         error("could not find object at offset %"PRIuMAX" "
1554                               "in pack %s", (uintmax_t)obj_offset, p->pack_name);
1555                         type = OBJ_BAD;
1556                         goto out;
1557                 }
1558
1559                 *oi->disk_sizep = pack_pos_to_offset(p, pos + 1) - obj_offset;
1560         }
1561
1562         if (oi->typep || oi->type_name) {
1563                 enum object_type ptot;
1564                 ptot = packed_to_object_type(r, p, obj_offset,
1565                                              type, &w_curs, curpos);
1566                 if (oi->typep)
1567                         *oi->typep = ptot;
1568                 if (oi->type_name) {
1569                         const char *tn = type_name(ptot);
1570                         if (tn)
1571                                 strbuf_addstr(oi->type_name, tn);
1572                 }
1573                 if (ptot < 0) {
1574                         type = OBJ_BAD;
1575                         goto out;
1576                 }
1577         }
1578
1579         if (oi->delta_base_oid) {
1580                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1581                         if (get_delta_base_oid(p, &w_curs, curpos,
1582                                                oi->delta_base_oid,
1583                                                type, obj_offset) < 0) {
1584                                 type = OBJ_BAD;
1585                                 goto out;
1586                         }
1587                 } else
1588                         oidclr(oi->delta_base_oid);
1589         }
1590
1591         oi->whence = in_delta_base_cache(p, obj_offset) ? OI_DBCACHED :
1592                                                           OI_PACKED;
1593
1594 out:
1595         unuse_pack(&w_curs);
1596         return type;
1597 }
1598
1599 static void *unpack_compressed_entry(struct packed_git *p,
1600                                     struct pack_window **w_curs,
1601                                     off_t curpos,
1602                                     unsigned long size)
1603 {
1604         int st;
1605         git_zstream stream;
1606         unsigned char *buffer, *in;
1607
1608         buffer = xmallocz_gently(size);
1609         if (!buffer)
1610                 return NULL;
1611         memset(&stream, 0, sizeof(stream));
1612         stream.next_out = buffer;
1613         stream.avail_out = size + 1;
1614
1615         git_inflate_init(&stream);
1616         do {
1617                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1618                 stream.next_in = in;
1619                 /*
1620                  * Note: we must ensure the window section returned by
1621                  * use_pack() will be available throughout git_inflate()'s
1622                  * unlocked execution. Please refer to the comment at
1623                  * get_size_from_delta() to see how this is done.
1624                  */
1625                 obj_read_unlock();
1626                 st = git_inflate(&stream, Z_FINISH);
1627                 obj_read_lock();
1628                 if (!stream.avail_out)
1629                         break; /* the payload is larger than it should be */
1630                 curpos += stream.next_in - in;
1631         } while (st == Z_OK || st == Z_BUF_ERROR);
1632         git_inflate_end(&stream);
1633         if ((st != Z_STREAM_END) || stream.total_out != size) {
1634                 free(buffer);
1635                 return NULL;
1636         }
1637
1638         /* versions of zlib can clobber unconsumed portion of outbuf */
1639         buffer[size] = '\0';
1640
1641         return buffer;
1642 }
1643
1644 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
1645 {
1646         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
1647         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
1648                          p->pack_name, (uintmax_t)obj_offset);
1649 }
1650
1651 int do_check_packed_object_crc;
1652
1653 #define UNPACK_ENTRY_STACK_PREALLOC 64
1654 struct unpack_entry_stack_ent {
1655         off_t obj_offset;
1656         off_t curpos;
1657         unsigned long size;
1658 };
1659
1660 static void *read_object(struct repository *r,
1661                          const struct object_id *oid,
1662                          enum object_type *type,
1663                          unsigned long *size)
1664 {
1665         struct object_info oi = OBJECT_INFO_INIT;
1666         void *content;
1667         oi.typep = type;
1668         oi.sizep = size;
1669         oi.contentp = &content;
1670
1671         if (oid_object_info_extended(r, oid, &oi, 0) < 0)
1672                 return NULL;
1673         return content;
1674 }
1675
1676 void *unpack_entry(struct repository *r, struct packed_git *p, off_t obj_offset,
1677                    enum object_type *final_type, unsigned long *final_size)
1678 {
1679         struct pack_window *w_curs = NULL;
1680         off_t curpos = obj_offset;
1681         void *data = NULL;
1682         unsigned long size;
1683         enum object_type type;
1684         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
1685         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
1686         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
1687         int base_from_cache = 0;
1688
1689         write_pack_access_log(p, obj_offset);
1690
1691         /* PHASE 1: drill down to the innermost base object */
1692         for (;;) {
1693                 off_t base_offset;
1694                 int i;
1695                 struct delta_base_cache_entry *ent;
1696
1697                 ent = get_delta_base_cache_entry(p, curpos);
1698                 if (ent) {
1699                         type = ent->type;
1700                         data = ent->data;
1701                         size = ent->size;
1702                         detach_delta_base_cache_entry(ent);
1703                         base_from_cache = 1;
1704                         break;
1705                 }
1706
1707                 if (do_check_packed_object_crc && p->index_version > 1) {
1708                         uint32_t pack_pos, index_pos;
1709                         off_t len;
1710
1711                         if (offset_to_pack_pos(p, obj_offset, &pack_pos) < 0) {
1712                                 error("could not find object at offset %"PRIuMAX" in pack %s",
1713                                       (uintmax_t)obj_offset, p->pack_name);
1714                                 data = NULL;
1715                                 goto out;
1716                         }
1717
1718                         len = pack_pos_to_offset(p, pack_pos + 1) - obj_offset;
1719                         index_pos = pack_pos_to_index(p, pack_pos);
1720                         if (check_pack_crc(p, &w_curs, obj_offset, len, index_pos)) {
1721                                 struct object_id oid;
1722                                 nth_packed_object_id(&oid, p, index_pos);
1723                                 error("bad packed object CRC for %s",
1724                                       oid_to_hex(&oid));
1725                                 mark_bad_packed_object(p, oid.hash);
1726                                 data = NULL;
1727                                 goto out;
1728                         }
1729                 }
1730
1731                 type = unpack_object_header(p, &w_curs, &curpos, &size);
1732                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
1733                         break;
1734
1735                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
1736                 if (!base_offset) {
1737                         error("failed to validate delta base reference "
1738                               "at offset %"PRIuMAX" from %s",
1739                               (uintmax_t)curpos, p->pack_name);
1740                         /* bail to phase 2, in hopes of recovery */
1741                         data = NULL;
1742                         break;
1743                 }
1744
1745                 /* push object, proceed to base */
1746                 if (delta_stack_nr >= delta_stack_alloc
1747                     && delta_stack == small_delta_stack) {
1748                         delta_stack_alloc = alloc_nr(delta_stack_nr);
1749                         ALLOC_ARRAY(delta_stack, delta_stack_alloc);
1750                         COPY_ARRAY(delta_stack, small_delta_stack,
1751                                    delta_stack_nr);
1752                 } else {
1753                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
1754                 }
1755                 i = delta_stack_nr++;
1756                 delta_stack[i].obj_offset = obj_offset;
1757                 delta_stack[i].curpos = curpos;
1758                 delta_stack[i].size = size;
1759
1760                 curpos = obj_offset = base_offset;
1761         }
1762
1763         /* PHASE 2: handle the base */
1764         switch (type) {
1765         case OBJ_OFS_DELTA:
1766         case OBJ_REF_DELTA:
1767                 if (data)
1768                         BUG("unpack_entry: left loop at a valid delta");
1769                 break;
1770         case OBJ_COMMIT:
1771         case OBJ_TREE:
1772         case OBJ_BLOB:
1773         case OBJ_TAG:
1774                 if (!base_from_cache)
1775                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
1776                 break;
1777         default:
1778                 data = NULL;
1779                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1780                       type, (uintmax_t)obj_offset, p->pack_name);
1781         }
1782
1783         /* PHASE 3: apply deltas in order */
1784
1785         /* invariants:
1786          *   'data' holds the base data, or NULL if there was corruption
1787          */
1788         while (delta_stack_nr) {
1789                 void *delta_data;
1790                 void *base = data;
1791                 void *external_base = NULL;
1792                 unsigned long delta_size, base_size = size;
1793                 int i;
1794                 off_t base_obj_offset = obj_offset;
1795
1796                 data = NULL;
1797
1798                 if (!base) {
1799                         /*
1800                          * We're probably in deep shit, but let's try to fetch
1801                          * the required base anyway from another pack or loose.
1802                          * This is costly but should happen only in the presence
1803                          * of a corrupted pack, and is better than failing outright.
1804                          */
1805                         uint32_t pos;
1806                         struct object_id base_oid;
1807                         if (!(offset_to_pack_pos(p, obj_offset, &pos))) {
1808                                 nth_packed_object_id(&base_oid, p,
1809                                                      pack_pos_to_index(p, pos));
1810                                 error("failed to read delta base object %s"
1811                                       " at offset %"PRIuMAX" from %s",
1812                                       oid_to_hex(&base_oid), (uintmax_t)obj_offset,
1813                                       p->pack_name);
1814                                 mark_bad_packed_object(p, base_oid.hash);
1815                                 base = read_object(r, &base_oid, &type, &base_size);
1816                                 external_base = base;
1817                         }
1818                 }
1819
1820                 i = --delta_stack_nr;
1821                 obj_offset = delta_stack[i].obj_offset;
1822                 curpos = delta_stack[i].curpos;
1823                 delta_size = delta_stack[i].size;
1824
1825                 if (!base)
1826                         continue;
1827
1828                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
1829
1830                 if (!delta_data) {
1831                         error("failed to unpack compressed delta "
1832                               "at offset %"PRIuMAX" from %s",
1833                               (uintmax_t)curpos, p->pack_name);
1834                         data = NULL;
1835                 } else {
1836                         data = patch_delta(base, base_size, delta_data,
1837                                            delta_size, &size);
1838
1839                         /*
1840                          * We could not apply the delta; warn the user, but
1841                          * keep going. Our failure will be noticed either in
1842                          * the next iteration of the loop, or if this is the
1843                          * final delta, in the caller when we return NULL.
1844                          * Those code paths will take care of making a more
1845                          * explicit warning and retrying with another copy of
1846                          * the object.
1847                          */
1848                         if (!data)
1849                                 error("failed to apply delta");
1850                 }
1851
1852                 /*
1853                  * We delay adding `base` to the cache until the end of the loop
1854                  * because unpack_compressed_entry() momentarily releases the
1855                  * obj_read_mutex, giving another thread the chance to access
1856                  * the cache. Therefore, if `base` was already there, this other
1857                  * thread could free() it (e.g. to make space for another entry)
1858                  * before we are done using it.
1859                  */
1860                 if (!external_base)
1861                         add_delta_base_cache(p, base_obj_offset, base, base_size, type);
1862
1863                 free(delta_data);
1864                 free(external_base);
1865         }
1866
1867         if (final_type)
1868                 *final_type = type;
1869         if (final_size)
1870                 *final_size = size;
1871
1872 out:
1873         unuse_pack(&w_curs);
1874
1875         if (delta_stack != small_delta_stack)
1876                 free(delta_stack);
1877
1878         return data;
1879 }
1880
1881 int bsearch_pack(const struct object_id *oid, const struct packed_git *p, uint32_t *result)
1882 {
1883         const unsigned char *index_fanout = p->index_data;
1884         const unsigned char *index_lookup;
1885         const unsigned int hashsz = the_hash_algo->rawsz;
1886         int index_lookup_width;
1887
1888         if (!index_fanout)
1889                 BUG("bsearch_pack called without a valid pack-index");
1890
1891         index_lookup = index_fanout + 4 * 256;
1892         if (p->index_version == 1) {
1893                 index_lookup_width = hashsz + 4;
1894                 index_lookup += 4;
1895         } else {
1896                 index_lookup_width = hashsz;
1897                 index_fanout += 8;
1898                 index_lookup += 8;
1899         }
1900
1901         return bsearch_hash(oid->hash, (const uint32_t*)index_fanout,
1902                             index_lookup, index_lookup_width, result);
1903 }
1904
1905 int nth_packed_object_id(struct object_id *oid,
1906                          struct packed_git *p,
1907                          uint32_t n)
1908 {
1909         const unsigned char *index = p->index_data;
1910         const unsigned int hashsz = the_hash_algo->rawsz;
1911         if (!index) {
1912                 if (open_pack_index(p))
1913                         return -1;
1914                 index = p->index_data;
1915         }
1916         if (n >= p->num_objects)
1917                 return -1;
1918         index += 4 * 256;
1919         if (p->index_version == 1) {
1920                 oidread(oid, index + (hashsz + 4) * n + 4);
1921         } else {
1922                 index += 8;
1923                 oidread(oid, index + hashsz * n);
1924         }
1925         return 0;
1926 }
1927
1928 void check_pack_index_ptr(const struct packed_git *p, const void *vptr)
1929 {
1930         const unsigned char *ptr = vptr;
1931         const unsigned char *start = p->index_data;
1932         const unsigned char *end = start + p->index_size;
1933         if (ptr < start)
1934                 die(_("offset before start of pack index for %s (corrupt index?)"),
1935                     p->pack_name);
1936         /* No need to check for underflow; .idx files must be at least 8 bytes */
1937         if (ptr >= end - 8)
1938                 die(_("offset beyond end of pack index for %s (truncated index?)"),
1939                     p->pack_name);
1940 }
1941
1942 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
1943 {
1944         const unsigned char *index = p->index_data;
1945         const unsigned int hashsz = the_hash_algo->rawsz;
1946         index += 4 * 256;
1947         if (p->index_version == 1) {
1948                 return ntohl(*((uint32_t *)(index + (hashsz + 4) * (size_t)n)));
1949         } else {
1950                 uint32_t off;
1951                 index += 8 + (size_t)p->num_objects * (hashsz + 4);
1952                 off = ntohl(*((uint32_t *)(index + 4 * n)));
1953                 if (!(off & 0x80000000))
1954                         return off;
1955                 index += (size_t)p->num_objects * 4 + (off & 0x7fffffff) * 8;
1956                 check_pack_index_ptr(p, index);
1957                 return get_be64(index);
1958         }
1959 }
1960
1961 off_t find_pack_entry_one(const unsigned char *sha1,
1962                                   struct packed_git *p)
1963 {
1964         const unsigned char *index = p->index_data;
1965         struct object_id oid;
1966         uint32_t result;
1967
1968         if (!index) {
1969                 if (open_pack_index(p))
1970                         return 0;
1971         }
1972
1973         hashcpy(oid.hash, sha1);
1974         if (bsearch_pack(&oid, p, &result))
1975                 return nth_packed_object_offset(p, result);
1976         return 0;
1977 }
1978
1979 int is_pack_valid(struct packed_git *p)
1980 {
1981         /* An already open pack is known to be valid. */
1982         if (p->pack_fd != -1)
1983                 return 1;
1984
1985         /* If the pack has one window completely covering the
1986          * file size, the pack is known to be valid even if
1987          * the descriptor is not currently open.
1988          */
1989         if (p->windows) {
1990                 struct pack_window *w = p->windows;
1991
1992                 if (!w->offset && w->len == p->pack_size)
1993                         return 1;
1994         }
1995
1996         /* Force the pack to open to prove its valid. */
1997         return !open_packed_git(p);
1998 }
1999
2000 struct packed_git *find_sha1_pack(const unsigned char *sha1,
2001                                   struct packed_git *packs)
2002 {
2003         struct packed_git *p;
2004
2005         for (p = packs; p; p = p->next) {
2006                 if (find_pack_entry_one(sha1, p))
2007                         return p;
2008         }
2009         return NULL;
2010
2011 }
2012
2013 static int fill_pack_entry(const struct object_id *oid,
2014                            struct pack_entry *e,
2015                            struct packed_git *p)
2016 {
2017         off_t offset;
2018
2019         if (p->num_bad_objects) {
2020                 unsigned i;
2021                 for (i = 0; i < p->num_bad_objects; i++)
2022                         if (hasheq(oid->hash,
2023                                    p->bad_object_sha1 + the_hash_algo->rawsz * i))
2024                                 return 0;
2025         }
2026
2027         offset = find_pack_entry_one(oid->hash, p);
2028         if (!offset)
2029                 return 0;
2030
2031         /*
2032          * We are about to tell the caller where they can locate the
2033          * requested object.  We better make sure the packfile is
2034          * still here and can be accessed before supplying that
2035          * answer, as it may have been deleted since the index was
2036          * loaded!
2037          */
2038         if (!is_pack_valid(p))
2039                 return 0;
2040         e->offset = offset;
2041         e->p = p;
2042         return 1;
2043 }
2044
2045 int find_pack_entry(struct repository *r, const struct object_id *oid, struct pack_entry *e)
2046 {
2047         struct list_head *pos;
2048         struct multi_pack_index *m;
2049
2050         prepare_packed_git(r);
2051         if (!r->objects->packed_git && !r->objects->multi_pack_index)
2052                 return 0;
2053
2054         for (m = r->objects->multi_pack_index; m; m = m->next) {
2055                 if (fill_midx_entry(r, oid, e, m))
2056                         return 1;
2057         }
2058
2059         list_for_each(pos, &r->objects->packed_git_mru) {
2060                 struct packed_git *p = list_entry(pos, struct packed_git, mru);
2061                 if (!p->multi_pack_index && fill_pack_entry(oid, e, p)) {
2062                         list_move(&p->mru, &r->objects->packed_git_mru);
2063                         return 1;
2064                 }
2065         }
2066         return 0;
2067 }
2068
2069 int has_object_pack(const struct object_id *oid)
2070 {
2071         struct pack_entry e;
2072         return find_pack_entry(the_repository, oid, &e);
2073 }
2074
2075 int has_pack_index(const unsigned char *sha1)
2076 {
2077         struct stat st;
2078         if (stat(sha1_pack_index_name(sha1), &st))
2079                 return 0;
2080         return 1;
2081 }
2082
2083 int for_each_object_in_pack(struct packed_git *p,
2084                             each_packed_object_fn cb, void *data,
2085                             enum for_each_object_flags flags)
2086 {
2087         uint32_t i;
2088         int r = 0;
2089
2090         if (flags & FOR_EACH_OBJECT_PACK_ORDER) {
2091                 if (load_pack_revindex(p))
2092                         return -1;
2093         }
2094
2095         for (i = 0; i < p->num_objects; i++) {
2096                 uint32_t index_pos;
2097                 struct object_id oid;
2098
2099                 /*
2100                  * We are iterating "i" from 0 up to num_objects, but its
2101                  * meaning may be different, depending on the requested output
2102                  * order:
2103                  *
2104                  *   - in object-name order, it is the same as the index order
2105                  *     used by nth_packed_object_id(), so we can pass it
2106                  *     directly
2107                  *
2108                  *   - in pack-order, it is pack position, which we must
2109                  *     convert to an index position in order to get the oid.
2110                  */
2111                 if (flags & FOR_EACH_OBJECT_PACK_ORDER)
2112                         index_pos = pack_pos_to_index(p, i);
2113                 else
2114                         index_pos = i;
2115
2116                 if (nth_packed_object_id(&oid, p, index_pos) < 0)
2117                         return error("unable to get sha1 of object %u in %s",
2118                                      index_pos, p->pack_name);
2119
2120                 r = cb(&oid, p, index_pos, data);
2121                 if (r)
2122                         break;
2123         }
2124         return r;
2125 }
2126
2127 int for_each_packed_object(each_packed_object_fn cb, void *data,
2128                            enum for_each_object_flags flags)
2129 {
2130         struct packed_git *p;
2131         int r = 0;
2132         int pack_errors = 0;
2133
2134         prepare_packed_git(the_repository);
2135         for (p = get_all_packs(the_repository); p; p = p->next) {
2136                 if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
2137                         continue;
2138                 if ((flags & FOR_EACH_OBJECT_PROMISOR_ONLY) &&
2139                     !p->pack_promisor)
2140                         continue;
2141                 if (open_pack_index(p)) {
2142                         pack_errors = 1;
2143                         continue;
2144                 }
2145                 r = for_each_object_in_pack(p, cb, data, flags);
2146                 if (r)
2147                         break;
2148         }
2149         return r ? r : pack_errors;
2150 }
2151
2152 static int add_promisor_object(const struct object_id *oid,
2153                                struct packed_git *pack,
2154                                uint32_t pos,
2155                                void *set_)
2156 {
2157         struct oidset *set = set_;
2158         struct object *obj = parse_object(the_repository, oid);
2159         if (!obj)
2160                 return 1;
2161
2162         oidset_insert(set, oid);
2163
2164         /*
2165          * If this is a tree, commit, or tag, the objects it refers
2166          * to are also promisor objects. (Blobs refer to no objects->)
2167          */
2168         if (obj->type == OBJ_TREE) {
2169                 struct tree *tree = (struct tree *)obj;
2170                 struct tree_desc desc;
2171                 struct name_entry entry;
2172                 if (init_tree_desc_gently(&desc, tree->buffer, tree->size))
2173                         /*
2174                          * Error messages are given when packs are
2175                          * verified, so do not print any here.
2176                          */
2177                         return 0;
2178                 while (tree_entry_gently(&desc, &entry))
2179                         oidset_insert(set, &entry.oid);
2180         } else if (obj->type == OBJ_COMMIT) {
2181                 struct commit *commit = (struct commit *) obj;
2182                 struct commit_list *parents = commit->parents;
2183
2184                 oidset_insert(set, get_commit_tree_oid(commit));
2185                 for (; parents; parents = parents->next)
2186                         oidset_insert(set, &parents->item->object.oid);
2187         } else if (obj->type == OBJ_TAG) {
2188                 struct tag *tag = (struct tag *) obj;
2189                 oidset_insert(set, get_tagged_oid(tag));
2190         }
2191         return 0;
2192 }
2193
2194 int is_promisor_object(const struct object_id *oid)
2195 {
2196         static struct oidset promisor_objects;
2197         static int promisor_objects_prepared;
2198
2199         if (!promisor_objects_prepared) {
2200                 if (has_promisor_remote()) {
2201                         for_each_packed_object(add_promisor_object,
2202                                                &promisor_objects,
2203                                                FOR_EACH_OBJECT_PROMISOR_ONLY);
2204                 }
2205                 promisor_objects_prepared = 1;
2206         }
2207         return oidset_contains(&promisor_objects, oid);
2208 }